Table of Contents

• 1  Loading and preparing data
  • 1.1  Loading original participant data (NZ)
    • 1.1.1  Melt dataframe to long format
  • 1.2  Loading replication participant data (US)
  • 1.3  Loading 2nd replication data (US, with reading measures)
    • 1.3.1  Compute reading subscales
  • 1.4  Retrieving vectors for words and dimension word pairs
• 2  Ranking color-semantic associations in word embeddings
  • 2.1  Using single dimension words
  • 2.2  Using dimension axes (word pair contrasts), with nearest neighbor (cosine) method
• 3  Creating datasets for statistical models
  • 3.1  Merging data and predictors
    • 3.1.1  Merge data
    • 3.1.2  Add word frequency (Van Paridon & Thompson, 2020)
    • 3.1.3  Add concreteness (Brysbaert et al., 2014)
    • 3.1.4  Add Small World of Words associations (De Deyne et al., 2018)
    • 3.1.5  Add cosine distances (Mikolov et al., 2013)
    • 3.1.6  Common Crawl
    • 3.1.7  Subtitles
  • 3.2  COCA
  • 3.3  Filtered COCA
  • 3.4  No neighbors COCA
  • 3.5  No names COCA
  • 3.6  Correlations between predictors
    • 3.6.1  Correlations between predictors in original data (NZ)
    • 3.6.2  Correlations between predictors in 2nd replication data (US)
    • 3.6.3  Correlations between predictors in full dataset
  • 3.7  Standardize predictors and write to file
• 4  Nameability of color-dimension associations
  • • 4.0.1  Exporting names generated by participants for use in training corpus filtering
    • 4.0.2  Correlating COCA-fiction cosine similarities to nameability measures
    • 4.0.3  Correlating group-averaged human ratings to nameability measure differentials.
    • 4.0.4  Correlation of group-averaged split-inverse ratings with nameability measures
• 5  Extracting non-color nearest neighbors for each dimension
• 6  More figures
  • 6.1  Mean color ratings on each dimension
  • 6.2  Scatterplot with connected points
• 7  Convert notebook to html

%matplotlib inline
%config InlineBackend.figure_format='retina'

from IPython.display import display, display_markdown

import warnings
warnings.simplefilter(action='ignore', category=FutureWarning)

import subprocess as sp
import numpy as np
import pandas as pd
import seaborn as sns
import pymc3 as pm
import arviz as az
import bambi
import copy
import warnings

import matplotlib.pyplot as plt
plt.rcParams['figure.figsize'] = [10, 8]
plt.rcParams['figure.dpi'] = 300

from scipy.stats import pearsonr, spearmanr
from itertools import chain
from collections import Counter

from subs2vec.utensils import log_timer
from subs2vec.vecs import Vectors
from subs2vec.neighbors import compute_nn

def display_md(md, **kwargs):
    return display_markdown(md, raw=True, **kwargs)

def convert_notebook(title, output='html'):
    convert = sp.run(f'jupyter nbconvert {title}.ipynb --to {output} --output {title}.{output}'.split(' '))
    if convert.returncode == 0:
        display_md(f'Jupyter notebook `{title}` converted successfully.')
    else:
        display_md(f'Error: encountered problem converting Jupyter notebook `{title}`')

def download(fname):
    dl = sp.run(f'wget {fname}'.split(' '))
    if dl.returncode == 0:
        display_md(f'Download of `{fname}` succesful.')
    else:
        display_md(f'Download of `{fname}` failed.')
        
@log_timer
def filter_vecs(vecs, filter_words):
    filtered_vecs = copy.deepcopy(vecs)
    filtered_vecs.vectors = filtered_vecs.vectors[np.isin(filtered_vecs.words, filter_words)]
    filtered_vecs.words = filtered_vecs.words[np.isin(filtered_vecs.words, filter_words)]
    filtered_vecs.n = len(filtered_vecs.words)
    display_md(f'Filtered {vecs.n} vectors, {filtered_vecs.n} remaining.')
    return filtered_vecs

def norm(x):
    return x / np.linalg.norm(x, 2)

sns.set(style='whitegrid')
pd.options.mode.chained_assignment = None

Loading and preparing data

Loading original participant data (NZ)

df = pd.read_csv('data/saysani_data.tsv', sep='\t')
display(df)

	participant	white	red	orange	yellow	green	blue	purple	brown	black	dimension	group	pp_id
0	1	1	7	7	5	1	1	1	3	7	cold-hot	sighted	sighted_1
1	1	7	1	4	2	3	3	6	6	7	ripe-unripe	sighted	sighted_1
2	1	1	5	6	7	4	2	3	7	6	new-old	sighted	sighted_1
3	1	1	7	2	1	4	2	3	5	7	submissive-aggressive	sighted	sighted_1
4	1	1	7	6	1	2	2	5	3	5	selfless-jealous	sighted	sighted_1
...	...	...	...	...	...	...	...	...	...	...	...	...	...
539	12	1	2	1	2	2	3	2	3	4	soft-hard	blind	blind_12
540	12	4	3	3	4	2	2	3	2	5	light-heavy	blind	blind_12
541	12	2	4	2	1	2	1	2	3	2	relaxed-tense	blind	blind_12
542	12	4	2	1	1	1	3	2	3	5	alive-dead	blind	blind_12
543	12	6	7	4	3	4	4	1	2	5	fast-slow	blind	blind_12

544 rows × 13 columns

Melt dataframe to long format

# these are the colors in the data
colors = ['white', 'red', 'orange', 'yellow', 'green', 'blue', 'purple', 'brown', 'black']

# melt
df_orig = df.melt(
    id_vars=['group', 'dimension', 'pp_id'],
    value_vars=colors,
    var_name='color',
    value_name='rating',
)

# pull out dimension words
dimension_labels = df_orig['dimension'].unique()
dimension_pairs = [pair.split('-') for pair in dimension_labels]
dimensions = list(chain(*dimension_pairs))

# add experiment and self vs. other variables for when we add the replication experiment later
df_orig['experiment'] = 'original'
df_orig['self_vs_other'] = 'self'

display(df_orig)

	group	dimension	pp_id	color	rating	experiment	self_vs_other
0	sighted	cold-hot	sighted_1	white	1	original	self
1	sighted	ripe-unripe	sighted_1	white	7	original	self
2	sighted	new-old	sighted_1	white	1	original	self
3	sighted	submissive-aggressive	sighted_1	white	1	original	self
4	sighted	selfless-jealous	sighted_1	white	1	original	self
...	...	...	...	...	...	...	...
4891	blind	soft-hard	blind_12	black	4	original	self
4892	blind	light-heavy	blind_12	black	5	original	self
4893	blind	relaxed-tense	blind_12	black	2	original	self
4894	blind	alive-dead	blind_12	black	5	original	self
4895	blind	fast-slow	blind_12	black	5	original	self

4896 rows × 7 columns

Loading replication participant data (US)

df_rep = pd.read_csv('data/replication1_data.csv')

# little bit of data munging, drop test participant and catch trials
df_rep = df_rep[(df_rep['pp_id'] != 3) & (df_rep['question_type'] != 'catch')]
df_rep = df_rep.drop(columns=['question_type', 'prompt_pre_1'])

# melt to long format
df_rep = df_rep.melt(
    id_vars=['dimension', 'color', 'pp_id'],
    value_vars=['value', 'others_choice'],
    var_name='self_vs_other',
    value_name='rating',
)

# more data munging
df_rep['pp_id'] = 'sighted_' + df_rep['pp_id'].astype(str)
df_rep['self_vs_other'] = df_rep['self_vs_other'].replace({'value': 'self', 'others_choice': 'other'})
df_rep['group'] = 'sighted'
df_rep['experiment'] = 'replication_1'

# there is a weird typo in one of the dimensions (?), so let's correct that here as well
df_rep['dimension'] = df_rep['dimension'].replace({'like-dis...like': 'like-dislike'})

display(df_rep)

	dimension	color	pp_id	self_vs_other	rating	group	experiment
0	clean-dirty	yellow	sighted_69819	self	5	sighted	replication_1
1	soft-hard	yellow	sighted_69819	self	2	sighted	replication_1
2	ripe-unripe	yellow	sighted_69819	self	1	sighted	replication_1
3	selfless-jealous	yellow	sighted_69819	self	5	sighted	replication_1
4	high-low	yellow	sighted_69819	self	1	sighted	replication_1
...	...	...	...	...	...	...	...
9567	like-dislike	orange	sighted_69785	other	4	sighted	replication_1
9568	new-old	orange	sighted_69785	other	4	sighted	replication_1
9569	clean-dirty	orange	sighted_69785	other	5	sighted	replication_1
9570	relaxed-tense	orange	sighted_69785	other	5	sighted	replication_1
9571	active-passive	orange	sighted_69785	other	3	sighted	replication_1

9572 rows × 7 columns

Loading 2nd replication data (US, with reading measures)

df_read = pd.read_csv('data/replication2_data_with_reading.csv').drop(columns=['Unnamed: 0', 'X'])

display(df_read)

	dimension	group	subj_id	color	value	question_type	others_choice	art	fiction	nonfiction	...	Q9_17	Q9_18	Q9_19	Q9_20	Q9_21	composite_read	upper_art	upper_fiction	upper_nonfiction	upper_read_motivation
0	cold-hot	replication-sighted	69212	brown	4	semantic_diff	4	3.0	0.0	1.0	...	2.0	2.0	2.0	2.0	2.0	2.000000	0.0	0.0	1.0	1.0
1	ripe-unripe	replication-sighted	69212	brown	7	semantic_diff	6	3.0	0.0	1.0	...	2.0	2.0	2.0	2.0	2.0	2.000000	0.0	0.0	1.0	1.0
2	new-old	replication-sighted	69212	brown	6	semantic_diff	6	3.0	0.0	1.0	...	2.0	2.0	2.0	2.0	2.0	2.000000	0.0	0.0	1.0	1.0
3	submissive-aggressive	replication-sighted	69212	brown	2	semantic_diff	2	3.0	0.0	1.0	...	2.0	2.0	2.0	2.0	2.0	2.000000	0.0	0.0	1.0	1.0
4	selfless-jealous	replication-sighted	69212	brown	5	semantic_diff	4	3.0	0.0	1.0	...	2.0	2.0	2.0	2.0	2.0	2.000000	0.0	0.0	1.0	1.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
14251	light-heavy	replication-sighted	68129	red	6	semantic_diff	5	11.0	0.0	1.0	...	1.0	0.0	-1.0	0.0	1.0	-0.555556	1.0	0.0	1.0	0.0
14252	relaxed-tense	replication-sighted	68129	red	6	semantic_diff	5	11.0	0.0	1.0	...	1.0	0.0	-1.0	0.0	1.0	-0.555556	1.0	0.0	1.0	0.0
14253	alive-dead	replication-sighted	68129	red	7	semantic_diff	6	11.0	0.0	1.0	...	1.0	0.0	-1.0	0.0	1.0	-0.555556	1.0	0.0	1.0	0.0
14254	fast-slow	replication-sighted	68129	red	1	semantic_diff	3	11.0	0.0	1.0	...	1.0	0.0	-1.0	0.0	1.0	-0.555556	1.0	0.0	1.0	0.0
14255	high-low	replication-sighted	68129	red	1	semantic_diff	2	11.0	0.0	1.0	...	1.0	0.0	-1.0	0.0	1.0	-0.555556	1.0	0.0	1.0	0.0

14256 rows × 36 columns

Compute reading subscales

df_read['reading_motivation'] = df_read.apply(
    lambda x: (0
    + x['Q9_1']
    + x['Q9_2']
    + x['Q9_3']
    + x['Q9_4']
    + x['Q9_5']
    + x['Q9_6']
    + x['Q9_7']
    + x['Q9_8']
    + x['Q9_9']
    + x['Q9_10']
    + x['Q9_11']
    + x['Q9_12']
    + x['Q9_13']
    + x['Q9_14'] * -1
    + x['Q9_15']
    + x['Q9_16']
    + x['Q9_17'] * -1
    + x['Q9_18']
    + x['Q9_19']
    + x['Q9_20']
    + x['Q9_21']) / 21,
axis=1)

df_read['reading_part_of_self'] = df_read.apply(
    lambda x: (0
    + x['Q9_2']
    + x['Q9_3']
    + x['Q9_4']
    + x['Q9_5']
    + x['Q9_6']
    + x['Q9_9']
    + x['Q9_10']
    + x['Q9_11']) / 8,
axis=1)

df_read['reading_efficacy'] = df_read.apply(
    lambda x: (0
    + x['Q9_1']
    + x['Q9_14'] * -1
    + x['Q9_16']
    + x['Q9_17'] * -1
    + x['Q9_19']
    + x['Q9_20']) / 6,
axis=1)

df_read['reading_recognition'] = df_read.apply(
    lambda x: (0
    + x['Q9_12']
    + x['Q9_13']
    + x['Q9_15']) / 3,
axis=1)

df_read['reading_other_realms'] = df_read.apply(
    lambda x: (0
    + x['Q9_7']
    + x['Q9_8']
    + x['Q9_18']
    + x['Q9_21']) / 4,
axis=1)

# rename participant id column to match earlier datasets
df_read = df_read.rename(columns={'subj_id': 'pp_id'})

# melt to long format
df_read = df_read.melt(
    id_vars=['dimension', 'color', 'pp_id', 'art', 'fiction', 'nonfiction', 'reading_motivation',
             'reading_part_of_self', 'reading_efficacy', 'reading_recognition', 'reading_other_realms'],
    value_vars=['value', 'others_choice'],
    var_name='self_vs_other',
    value_name='rating',
)

# more data munging
df_read['pp_id'] = 'sighted_' + df_read['pp_id'].astype(str)
df_read['self_vs_other'] = df_read['self_vs_other'].replace({'value': 'self', 'others_choice': 'other'})
df_read['group'] = 'sighted'
df_read['experiment'] = 'replication_2'

display(df_read)

	dimension	color	pp_id	art	fiction	nonfiction	reading_motivation	reading_part_of_self	reading_efficacy	reading_recognition	reading_other_realms	self_vs_other	rating	group	experiment
0	cold-hot	brown	sighted_69212	3.0	0.0	1.0	1.619048	2.000	0.666667	2.0	2.00	self	4	sighted	replication_2
1	ripe-unripe	brown	sighted_69212	3.0	0.0	1.0	1.619048	2.000	0.666667	2.0	2.00	self	7	sighted	replication_2
2	new-old	brown	sighted_69212	3.0	0.0	1.0	1.619048	2.000	0.666667	2.0	2.00	self	6	sighted	replication_2
3	submissive-aggressive	brown	sighted_69212	3.0	0.0	1.0	1.619048	2.000	0.666667	2.0	2.00	self	2	sighted	replication_2
4	selfless-jealous	brown	sighted_69212	3.0	0.0	1.0	1.619048	2.000	0.666667	2.0	2.00	self	5	sighted	replication_2
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
28507	light-heavy	red	sighted_68129	11.0	0.0	1.0	-0.571429	-0.375	-0.833333	-1.0	-0.25	other	5	sighted	replication_2
28508	relaxed-tense	red	sighted_68129	11.0	0.0	1.0	-0.571429	-0.375	-0.833333	-1.0	-0.25	other	5	sighted	replication_2
28509	alive-dead	red	sighted_68129	11.0	0.0	1.0	-0.571429	-0.375	-0.833333	-1.0	-0.25	other	6	sighted	replication_2
28510	fast-slow	red	sighted_68129	11.0	0.0	1.0	-0.571429	-0.375	-0.833333	-1.0	-0.25	other	3	sighted	replication_2
28511	high-low	red	sighted_68129	11.0	0.0	1.0	-0.571429	-0.375	-0.833333	-1.0	-0.25	other	2	sighted	replication_2

28512 rows × 15 columns

df_read.describe()

	art	fiction	nonfiction	reading_motivation	reading_part_of_self	reading_efficacy	reading_recognition	reading_other_realms	rating
count	27864.000000	27864.000000	27864.000000	27864.000000	27864.000000	27864.000000	27864.000000	27864.000000	28512.000000
mean	7.616279	0.593023	0.755814	-0.107973	-0.280523	0.203488	-0.616279	0.151163	3.693147
std	6.612596	0.854251	0.987569	0.646210	0.879559	0.646210	0.910077	0.759909	1.424941
min	-5.000000	0.000000	0.000000	-1.619048	-2.000000	-1.000000	-2.000000	-2.000000	1.000000
25%	3.000000	0.000000	0.000000	-0.571429	-1.000000	-0.333333	-1.333333	-0.500000	3.000000
50%	6.000000	0.000000	0.000000	-0.119048	-0.375000	0.166667	-0.666667	0.250000	4.000000
75%	10.000000	1.000000	1.000000	0.285714	0.250000	0.666667	0.000000	0.500000	5.000000
max	26.000000	4.000000	4.000000	1.619048	2.000000	2.000000	2.000000	2.000000	7.000000

corrs = df_read[['art', 'fiction', 'nonfiction', 'reading_motivation', 'reading_part_of_self',
                 'reading_efficacy', 'reading_recognition', 'reading_other_realms']].corr().round(2)
mask = np.zeros_like(corrs)
mask[np.triu_indices_from(mask)] = True
g = sns.heatmap(corrs, vmin=0, vmax=1, annot=True, mask=mask)
g.set_yticklabels(g.get_yticklabels(), rotation=0);

g = sns.histplot(x='art', data=df_read)

Retrieving vectors for words and dimension word pairs

vecs = Vectors('../embeddings/fic.en.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs = filter_vecs(vecs, np.array(colors + dimensions))
vecs_dict = vecs.as_dict()

color_vecs = filter_vecs(vecs, np.array(colors))
dimension_vecs = filter_vecs(vecs, np.array(dimensions))

dimension_pair_vecs = np.vstack([norm(vecs_dict[pair[0]] - vecs_dict[pair[1]]) for pair in dimension_pairs])

[INFO] loading vectors ../embeddings/fic.en.vec
[INFO] <function Vectors.__init__ at 0x13d7ac0d0> ran in 5.879 seconds

Filtered 200000 vectors, 43 remaining.

[INFO] <function filter_vecs at 0x14acad040> ran in 0.242 seconds
[INFO] <function Vectors.as_dict at 0x13d7ac310> ran in 0.000 seconds

Filtered 43 vectors, 9 remaining.

[INFO] <function filter_vecs at 0x14acad040> ran in 0.000 seconds

Filtered 43 vectors, 34 remaining.

[INFO] <function filter_vecs at 0x14acad040> ran in 0.000 seconds

Ranking color-semantic associations in word embeddings

Using single dimension words

dimension_neighbors = compute_nn(color_vecs, dimension_vecs.vectors, dimension_vecs.words, num_neighbors=9, whole_matrix=True)
dimension_neighbors = dimension_neighbors.drop(columns=[
    'neighbor -1',
    'neighbor -2',
    'neighbor -3',
    'neighbor -4',
    'neighbor -5',
    'neighbor -6',
    'neighbor -7',
    'neighbor -8',
    'neighbor -9'
]).rename(columns={'target': 'dimension'})
display(dimension_neighbors)

[INFO] <function Vectors.as_dict at 0x13d7ac310> ran in 0.000 seconds
[INFO] computing analogies using whole matrix additive method
[INFO] <function compute_nn at 0x13d7ac5e0> ran in 0.002 seconds

	dimension	neighbor 0	neighbor 1	neighbor 2	neighbor 3	neighbor 4	neighbor 5	neighbor 6	neighbor 7	neighbor 8
0	like	white	black	brown	yellow	orange	purple	blue	green	red
1	old	white	black	brown	yellow	orange	blue	red	green	purple
2	new	white	green	yellow	black	red	blue	purple	brown	orange
3	light	yellow	red	green	orange	blue	purple	white	brown	black
4	hard	brown	white	black	red	orange	purple	yellow	green	blue
5	dead	black	brown	white	red	green	yellow	purple	blue	orange
6	cold	blue	white	green	black	red	brown	purple	yellow	orange
7	happy	white	yellow	orange	red	brown	blue	green	purple	black
8	hot	red	yellow	black	purple	white	orange	blue	green	brown
9	heavy	black	brown	green	red	purple	yellow	blue	white	orange
10	fast	red	purple	white	blue	yellow	green	black	brown	orange
11	soft	brown	yellow	white	green	orange	blue	purple	red	black
12	clean	white	blue	black	yellow	brown	green	red	orange	purple
13	slow	red	yellow	brown	black	green	purple	white	orange	blue
14	angry	red	orange	purple	blue	white	yellow	black	brown	green
15	alive	green	black	orange	red	blue	yellow	brown	purple	white
16	sad	brown	purple	blue	white	green	black	red	yellow	orange
17	fresh	white	green	blue	yellow	brown	red	black	purple	orange
18	calm	blue	green	brown	white	purple	yellow	black	red	orange
19	dirty	white	brown	yellow	blue	black	red	green	orange	purple
20	dull	brown	green	yellow	blue	orange	red	purple	black	white
21	relaxed	brown	white	green	yellow	red	blue	orange	black	purple
22	jealous	red	purple	orange	black	green	blue	yellow	brown	white
23	tense	white	black	red	yellow	blue	brown	purple	green	orange
24	exciting	green	orange	black	brown	purple	blue	red	white	yellow
25	active	orange	black	green	yellow	red	white	brown	blue	purple
26	ripe	purple	orange	brown	green	yellow	blue	red	white	black
27	aggressive	black	brown	orange	white	yellow	red	purple	green	blue
28	stale	brown	white	yellow	black	red	orange	green	blue	purple
29	dislike	brown	green	black	blue	orange	purple	white	yellow	red
30	passive	white	black	brown	yellow	red	green	purple	orange	blue
31	selfless	blue	purple	white	black	green	red	orange	yellow	brown
32	submissive	white	brown	yellow	purple	black	blue	orange	green	red
33	unripe	orange	purple	yellow	brown	green	blue	red	white	black

Using dimension axes (word pair contrasts), with nearest neighbor (cosine) method

dimension_neighbors = compute_nn(color_vecs, dimension_pair_vecs, np.array(dimension_labels), num_neighbors=9, whole_matrix=True)
dimension_neighbors = dimension_neighbors.drop(columns=[
    'neighbor -1',
    'neighbor -2',
    'neighbor -3',
    'neighbor -4',
    'neighbor -5',
    'neighbor -6',
    'neighbor -7',
    'neighbor -8',
    'neighbor -9'
]).rename(columns={'target': 'dimension'})
display(dimension_neighbors)

[INFO] <function Vectors.as_dict at 0x13d7ac310> ran in 0.000 seconds
[INFO] computing analogies using whole matrix additive method
[INFO] <function compute_nn at 0x13d7ac5e0> ran in 0.001 seconds

	dimension	neighbor 0	neighbor 1	neighbor 2	neighbor 3	neighbor 4	neighbor 5	neighbor 6	neighbor 7	neighbor 8
0	cold-hot	blue	white	green	brown	black	purple	red	yellow	orange
1	ripe-unripe	green	white	brown	black	blue	red	purple	yellow	orange
2	new-old	green	purple	yellow	red	blue	black	orange	white	brown
3	submissive-aggressive	white	blue	purple	yellow	brown	green	black	orange	red
4	selfless-jealous	white	blue	purple	brown	green	yellow	black	orange	red
5	active-passive	orange	green	blue	red	black	yellow	purple	brown	white
6	like-dislike	white	black	yellow	red	orange	brown	purple	blue	green
7	clean-dirty	white	blue	black	purple	green	red	orange	yellow	brown
8	fresh-stale	green	blue	purple	red	white	yellow	black	orange	brown
9	calm-angry	green	blue	brown	white	black	yellow	purple	orange	red
10	happy-sad	white	orange	yellow	red	blue	green	black	brown	purple
11	exciting-dull	black	white	purple	orange	red	green	blue	yellow	brown
12	soft-hard	yellow	brown	blue	green	orange	purple	white	red	black
13	light-heavy	yellow	orange	blue	red	green	white	purple	brown	black
14	relaxed-tense	brown	green	orange	yellow	white	blue	red	purple	black
15	alive-dead	orange	blue	green	purple	yellow	red	brown	black	white
16	fast-slow	white	blue	purple	orange	green	red	black	yellow	brown

Creating datasets for statistical models

Merging data and predictors

Merge data

df_joint = pd.concat([df_orig, df_rep, df_read]).reset_index()
display(df_joint)

	index	group	dimension	pp_id	color	rating	experiment	self_vs_other	art	fiction	nonfiction	reading_motivation	reading_part_of_self	reading_efficacy	reading_recognition	reading_other_realms
0	0	sighted	cold-hot	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
1	1	sighted	ripe-unripe	sighted_1	white	7	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
2	2	sighted	new-old	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
3	3	sighted	submissive-aggressive	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
4	4	sighted	selfless-jealous	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
42975	28507	sighted	light-heavy	sighted_68129	red	5	replication_2	other	11.0	0.0	1.0	-0.571429	-0.375	-0.833333	-1.0	-0.25
42976	28508	sighted	relaxed-tense	sighted_68129	red	5	replication_2	other	11.0	0.0	1.0	-0.571429	-0.375	-0.833333	-1.0	-0.25
42977	28509	sighted	alive-dead	sighted_68129	red	6	replication_2	other	11.0	0.0	1.0	-0.571429	-0.375	-0.833333	-1.0	-0.25
42978	28510	sighted	fast-slow	sighted_68129	red	3	replication_2	other	11.0	0.0	1.0	-0.571429	-0.375	-0.833333	-1.0	-0.25
42979	28511	sighted	high-low	sighted_68129	red	2	replication_2	other	11.0	0.0	1.0	-0.571429	-0.375	-0.833333	-1.0	-0.25

42980 rows × 16 columns

Add word frequency (Van Paridon & Thompson, 2020)

freqs = pd.read_csv('../datasets/dedup.en.words.unigrams.tsv', sep='\t')  # not included in git repo
freqs['log_freq'] = np.log(freqs['unigram_freq'])
freqs = freqs.drop(columns='unigram_freq')
display(freqs.round(2))

	unigram	log_freq
0	the	17.10
1	you	17.06
2	i	17.04
3	to	16.78
4	a	16.59
...	...	...
2397976	tpar1	0.00
2397977	giacoia	0.00
2397978	ourcinders	0.00
2397979	tourret	0.00
2397980	iroki	0.00

2397981 rows × 2 columns

df_joint['word1'] = df_joint['dimension'].apply(lambda x: x.split('-')[0])
df_joint['word2'] = df_joint['dimension'].apply(lambda x: x.split('-')[1])
df_joint = df_joint.merge(freqs, left_on='word1', right_on='unigram', how='left')
df_joint = df_joint.merge(freqs, left_on='word2', right_on='unigram', how='left')
df_joint['frequency'] = df_joint['log_freq_x'] - df_joint['log_freq_y']
df_joint = df_joint.drop(columns=[
    'unigram_x',
    'unigram_y',
    'log_freq_x',
    'log_freq_y'
])
display(df_joint.round(2))

	index	group	dimension	pp_id	color	rating	experiment	self_vs_other	art	fiction	nonfiction	reading_motivation	reading_part_of_self	reading_efficacy	reading_recognition	reading_other_realms	word1	word2	frequency
0	0	sighted	cold-hot	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	cold	hot	-0.22
1	1	sighted	ripe-unripe	sighted_1	white	7	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	ripe	unripe	3.49
2	2	sighted	new-old	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	new	old	0.12
3	3	sighted	submissive-aggressive	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	submissive	aggressive	-2.35
4	4	sighted	selfless-jealous	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	selfless	jealous	-2.96
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
42975	28507	sighted	light-heavy	sighted_68129	red	5	replication_2	other	11.0	0.0	1.0	-0.57	-0.38	-0.83	-1.0	-0.25	light	heavy	1.24
42976	28508	sighted	relaxed-tense	sighted_68129	red	5	replication_2	other	11.0	0.0	1.0	-0.57	-0.38	-0.83	-1.0	-0.25	relaxed	tense	-0.23
42977	28509	sighted	alive-dead	sighted_68129	red	6	replication_2	other	11.0	0.0	1.0	-0.57	-0.38	-0.83	-1.0	-0.25	alive	dead	-0.90
42978	28510	sighted	fast-slow	sighted_68129	red	3	replication_2	other	11.0	0.0	1.0	-0.57	-0.38	-0.83	-1.0	-0.25	fast	slow	0.76
42979	28511	sighted	high-low	sighted_68129	red	2	replication_2	other	11.0	0.0	1.0	-0.57	-0.38	-0.83	-1.0	-0.25	high	low	1.24

42980 rows × 19 columns

Add concreteness (Brysbaert et al., 2014)

concreteness = pd.read_csv('../datasets/en-brysbaert-2014.tsv', sep='\t')  # not included in git repo
display(concreteness)

	word	concreteness
0	a	1.46
1	aardvark	4.68
2	aback	1.65
3	abacus	4.52
4	abandon	2.54
...	...	...
37053	zoologist	4.30
37054	zoology	3.37
37055	zoom	3.10
37056	zoophobia	2.04
37057	zucchini	4.87

37058 rows × 2 columns

df_joint = df_joint.merge(concreteness, left_on='word1', right_on='word', how='left')
df_joint = df_joint.merge(concreteness, left_on='word2', right_on='word', how='left')
df_joint['concreteness'] = df_joint['concreteness_x'] - df_joint['concreteness_y']
df_joint = df_joint.drop(columns=[
    'word_x',
    'word_y',
    'concreteness_x',
    'concreteness_y'
])
display(df_joint.round(2))

	index	group	dimension	pp_id	color	rating	experiment	self_vs_other	art	fiction	nonfiction	reading_motivation	reading_part_of_self	reading_efficacy	reading_recognition	reading_other_realms	word1	word2	frequency	concreteness
0	0	sighted	cold-hot	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	cold	hot	-0.22	-0.46
1	1	sighted	ripe-unripe	sighted_1	white	7	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	ripe	unripe	3.49	-0.01
2	2	sighted	new-old	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	new	old	0.12	0.09
3	3	sighted	submissive-aggressive	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	submissive	aggressive	-2.35	-0.82
4	4	sighted	selfless-jealous	sighted_1	white	1	original	self	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	selfless	jealous	-2.96	-0.56
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
42975	28507	sighted	light-heavy	sighted_68129	red	5	replication_2	other	11.0	0.0	1.0	-0.57	-0.38	-0.83	-1.0	-0.25	light	heavy	1.24	0.84
42976	28508	sighted	relaxed-tense	sighted_68129	red	5	replication_2	other	11.0	0.0	1.0	-0.57	-0.38	-0.83	-1.0	-0.25	relaxed	tense	-0.23	0.15
42977	28509	sighted	alive-dead	sighted_68129	red	6	replication_2	other	11.0	0.0	1.0	-0.57	-0.38	-0.83	-1.0	-0.25	alive	dead	-0.90	-0.93
42978	28510	sighted	fast-slow	sighted_68129	red	3	replication_2	other	11.0	0.0	1.0	-0.57	-0.38	-0.83	-1.0	-0.25	fast	slow	0.76	0.04
42979	28511	sighted	high-low	sighted_68129	red	2	replication_2	other	11.0	0.0	1.0	-0.57	-0.38	-0.83	-1.0	-0.25	high	low	1.24	0.12

42980 rows × 20 columns

Add Small World of Words associations (De Deyne et al., 2018)

swow = pd.read_csv('../datasets/SWOW-EN.R100.csv')  # not included in git repo
display(swow)

	Unnamed: 0	id	participantID	age	gender	nativeLanguage	country	education	created_at	cue	R1	R2	R3
0	1	29	3	33	Fe	United States	Australia	NaN	2011-08-12 02:19:38	although	nevertheless	yet	but
1	2	30	3	33	Fe	United States	Australia	NaN	2011-08-12 02:19:38	deal	no	cards	shake
2	3	31	3	33	Fe	United States	Australia	NaN	2011-08-12 02:19:38	music	notes	band	rhythm
3	4	32	3	33	Fe	United States	Australia	NaN	2011-08-12 02:19:38	inform	tell	rat on	NaN
4	5	33	3	33	Fe	United States	Australia	NaN	2011-08-12 02:19:38	way	path	via	method
...	...	...	...	...	...	...	...	...	...	...	...	...	...
1228195	1228196	1530300	132506	29	Ma	Canada	Australia	5.0	2018-08-10 01:56:27	strange	mask	weird	stranger
1228196	1228197	1530290	132506	29	Ma	Canada	Australia	5.0	2018-08-10 01:56:27	sunset	sea	sky	clause
1228197	1228198	1530291	132506	29	Ma	Canada	Australia	5.0	2018-08-10 01:56:27	useless	pitty	worthless	worth
1228198	1228199	1530284	132506	29	Ma	Canada	Australia	5.0	2018-08-10 01:56:27	volume	loud	music	key
1228199	1228200	1530288	132506	29	Ma	Canada	Australia	5.0	2018-08-10 01:56:27	whenever	who	where	always

1228200 rows × 13 columns

def add_swow(df, swow, colname):
    swow = pd.DataFrame(swow.groupby('cue')['resp'].value_counts()).rename(columns={'resp': 'n'})
    swow = swow.reset_index()
    df = df.merge(swow, left_on=['word1', 'color'], right_on=['cue', 'resp'], how='left')
    df = df.merge(swow, left_on=['word2', 'color'], right_on=['cue', 'resp'], how='left')
    df['n_x'] = df['n_x'].fillna(0)
    df['n_y'] = df['n_y'].fillna(0)
    df[colname] = df['n_x'] - df['n_y']
    df = df.drop(columns=[
        'cue_x',
        'cue_y',
        'resp_x',
        'resp_y',
        'n_x',
        'n_y',
    ])
    return df

swow = swow[swow['cue'].isin(dimensions)]
swow_NZ = swow[(swow['country'] == 'New Zealand')]  # select only NZ respondents
swow_US = swow[(swow['country'] == 'United States')]  # select only US respondents

# count only R1 (maximal discounting)
df_joint = add_swow(df_joint, swow.rename(columns={'R1': 'resp'}), 'swow_R1')
df_joint = add_swow(df_joint, swow_NZ.rename(columns={'R1': 'resp'}), 'swow_R1_NZ')  # US
df_joint = add_swow(df_joint, swow_US.rename(columns={'R1': 'resp'}), 'swow_R1_US')  # NZ

# count R1, R2, and R3 with equal weight (minimal discounting)
swow_all = swow.melt(
    id_vars=['id', 'participantID', 'created_at', 'cue'],
    value_vars=['R1', 'R2', 'R3'],
    value_name='resp',
)
df_joint = add_swow(df_joint, swow_all, 'swow_all')

# NZ
swow_all_NZ = swow_NZ.melt(
    id_vars=['id', 'participantID', 'created_at', 'cue'],
    value_vars=['R1', 'R2', 'R3'],
    value_name='resp',
)
df_joint = add_swow(df_joint, swow_all_NZ, 'swow_all_NZ')

# US
swow_all_US = swow_US.melt(
    id_vars=['id', 'participantID', 'created_at', 'cue'],
    value_vars=['R1', 'R2', 'R3'],
    value_name='resp',
)
df_joint = add_swow(df_joint, swow_all_US, 'swow_all_US')

display(df_joint)

	index	group	dimension	pp_id	color	rating	experiment	self_vs_other	art	fiction	...	word1	word2	frequency	concreteness	swow_R1	swow_R1_NZ	swow_R1_US	swow_all	swow_all_NZ	swow_all_US
0	0	sighted	cold-hot	sighted_1	white	1	original	self	NaN	NaN	...	cold	hot	-0.216432	-0.46	0.0	0.0	0.0	0.0	0.0	0.0
1	1	sighted	ripe-unripe	sighted_1	white	7	original	self	NaN	NaN	...	ripe	unripe	3.485549	-0.01	0.0	0.0	0.0	0.0	0.0	0.0
2	2	sighted	new-old	sighted_1	white	1	original	self	NaN	NaN	...	new	old	0.119068	0.09	0.0	0.0	0.0	0.0	0.0	0.0
3	3	sighted	submissive-aggressive	sighted_1	white	1	original	self	NaN	NaN	...	submissive	aggressive	-2.352148	-0.82	0.0	0.0	0.0	0.0	0.0	0.0
4	4	sighted	selfless-jealous	sighted_1	white	1	original	self	NaN	NaN	...	selfless	jealous	-2.955968	-0.56	0.0	0.0	0.0	0.0	0.0	0.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
42975	28507	sighted	light-heavy	sighted_68129	red	5	replication_2	other	11.0	0.0	...	light	heavy	1.240142	0.84	0.0	0.0	0.0	0.0	0.0	0.0
42976	28508	sighted	relaxed-tense	sighted_68129	red	5	replication_2	other	11.0	0.0	...	relaxed	tense	-0.229652	0.15	0.0	0.0	0.0	0.0	0.0	0.0
42977	28509	sighted	alive-dead	sighted_68129	red	6	replication_2	other	11.0	0.0	...	alive	dead	-0.904786	-0.93	-1.0	0.0	0.0	-1.0	0.0	0.0
42978	28510	sighted	fast-slow	sighted_68129	red	3	replication_2	other	11.0	0.0	...	fast	slow	0.763262	0.04	0.0	0.0	0.0	0.0	0.0	0.0
42979	28511	sighted	high-low	sighted_68129	red	2	replication_2	other	11.0	0.0	...	high	low	1.237676	0.12	0.0	0.0	0.0	0.0	0.0	0.0

42980 rows × 26 columns

# check how many participants gave green as a response to various cues (to use as an example in the paper)
counts = swow_all_US.groupby(['cue', 'resp']).count().reset_index()
display(counts[counts['resp'] == 'green'])

	cue	resp	id	participantID	created_at	variable
233	alive	green	1	1	1	1
508	clean	green	1	1	1	1
1108	exciting	green	1	1	1	1
1289	fresh	green	1	1	1	1
1456	hard	green	1	1	1	1
1706	jealous	green	20	20	20	20
1984	new	green	1	1	1	1
3010	unripe	green	18	18	18	18

display(df_joint.sort_values('swow_all'))

	index	group	dimension	pp_id	color	rating	experiment	self_vs_other	art	fiction	...	word1	word2	frequency	concreteness	swow_R1	swow_R1_NZ	swow_R1_US	swow_all	swow_all_NZ	swow_all_US
25920	11452	sighted	selfless-jealous	sighted_68676	green	2	replication_2	self	4.0	0.0	...	selfless	jealous	-2.955968	-0.56	-19.0	0.0	-10.0	-40.0	-1.0	-20.0
5138	242	sighted	selfless-jealous	sighted_68736	green	2	replication_1	self	NaN	NaN	...	selfless	jealous	-2.955968	-0.56	-19.0	0.0	-10.0	-40.0	-1.0	-20.0
37530	23062	sighted	selfless-jealous	sighted_67653	green	6	replication_2	other	10.0	4.0	...	selfless	jealous	-2.955968	-0.56	-19.0	0.0	-10.0	-40.0	-1.0	-20.0
25416	10948	sighted	selfless-jealous	sighted_69192	green	7	replication_2	self	9.0	1.0	...	selfless	jealous	-2.955968	-0.56	-19.0	0.0	-10.0	-40.0	-1.0	-20.0
16956	2488	sighted	selfless-jealous	sighted_68719	green	5	replication_2	self	3.0	0.0	...	selfless	jealous	-2.955968	-0.56	-19.0	0.0	-10.0	-40.0	-1.0	-20.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
466	466	blind	clean-dirty	blind_8	white	2	original	self	NaN	NaN	...	clean	dirty	0.600633	-1.16	0.0	0.0	0.0	8.0	0.0	7.0
14129	9233	sighted	clean-dirty	sighted_68738	white	1	replication_1	other	NaN	NaN	...	clean	dirty	0.600633	-1.16	0.0	0.0	0.0	8.0	0.0	7.0
33057	18589	sighted	light-heavy	sighted_68150	white	2	replication_2	other	9.0	0.0	...	light	heavy	1.240142	0.84	1.0	0.0	1.0	8.0	0.0	5.0
12790	7894	sighted	clean-dirty	sighted_68946	white	1	replication_1	other	NaN	NaN	...	clean	dirty	0.600633	-1.16	0.0	0.0	0.0	8.0	0.0	7.0
21537	7069	sighted	light-heavy	sighted_67884	white	2	replication_2	self	5.0	1.0	...	light	heavy	1.240142	0.84	1.0	0.0	1.0	8.0	0.0	5.0

42980 rows × 26 columns

(It looks like there very few responses from NZ, but a little more from US and elsewhere.)

Add cosine distances (Mikolov et al., 2013)

def get_cosine(x, vecs_dict):
    return np.dot(norm(vecs_dict.get(x['word2'], 0) - vecs_dict.get(x['word1'], 0)), vecs_dict.get(x['color'], 0))

Common Crawl

vecs = Vectors('../embeddings/cc.en.300.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()
df_joint['cosine_cc'] = df_joint.apply(lambda x: get_cosine(x, vecs_dict), axis=1)

[INFO] loading vectors ../embeddings/cc.en.300.vec
[INFO] <function Vectors.__init__ at 0x13d7ac0d0> ran in 5.757 seconds
[INFO] <function Vectors.as_dict at 0x13d7ac310> ran in 0.054 seconds

Subtitles

vecs = Vectors('../embeddings/subs.en.1e6.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()
df_joint['cosine_subs'] = df_joint.apply(lambda x: get_cosine(x, vecs_dict), axis=1)

[INFO] loading vectors ../embeddings/subs.en.1e6.vec
[INFO] <function Vectors.__init__ at 0x13d7ac0d0> ran in 6.154 seconds
[INFO] <function Vectors.as_dict at 0x13d7ac310> ran in 0.041 seconds

COCA

# academic
vecs = Vectors('../embeddings/acad.en.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()
df_joint['cosine_acad'] = df_joint.apply(lambda x: get_cosine(x, vecs_dict), axis=1)

# fiction
vecs = Vectors('../embeddings/fic.en.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()
df_joint['cosine_fic'] = df_joint.apply(lambda x: get_cosine(x, vecs_dict), axis=1)

# magazines
vecs = Vectors('../embeddings/mag.en.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()
df_joint['cosine_mag'] = df_joint.apply(lambda x: get_cosine(x, vecs_dict), axis=1)

# spoken
vecs = Vectors('../embeddings/spok.en.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()
df_joint['cosine_spok'] = df_joint.apply(lambda x: get_cosine(x, vecs_dict), axis=1)

# news
vecs = Vectors('../embeddings/news.en.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()
df_joint['cosine_news'] = df_joint.apply(lambda x: get_cosine(x, vecs_dict), axis=1)

display(df_joint.round(2))

[INFO] loading vectors ../embeddings/acad.en.vec
[INFO] <function Vectors.__init__ at 0x7fb0731c9d40> ran in 16.646 seconds
[INFO] <function Vectors.as_dict at 0x7fb0731c9f80> ran in 0.151 seconds
[INFO] loading vectors ../embeddings/fic.en.vec
[INFO] <function Vectors.__init__ at 0x7fb0731c9d40> ran in 16.292 seconds
[INFO] <function Vectors.as_dict at 0x7fb0731c9f80> ran in 0.131 seconds
[INFO] loading vectors ../embeddings/mag.en.vec
[INFO] <function Vectors.__init__ at 0x7fb0731c9d40> ran in 15.646 seconds
[INFO] <function Vectors.as_dict at 0x7fb0731c9f80> ran in 0.142 seconds
[INFO] loading vectors ../embeddings/spok.en.vec
[INFO] <function Vectors.__init__ at 0x7fb0731c9d40> ran in 15.418 seconds
[INFO] <function Vectors.as_dict at 0x7fb0731c9f80> ran in 0.136 seconds
[INFO] loading vectors ../embeddings/news.en.vec
[INFO] <function Vectors.__init__ at 0x7fb0731c9d40> ran in 16.070 seconds
[INFO] <function Vectors.as_dict at 0x7fb0731c9f80> ran in 0.155 seconds

	index	group	dimension	pp_id	color	rating	experiment	self_vs_other	art	fiction	...	swow_all	swow_all_NZ	swow_all_US	cosine_cc	cosine_subs	cosine_acad	cosine_fic	cosine_mag	cosine_spok	cosine_news
0	0	sighted	cold-hot	sighted_1	white	1	original	self	NaN	NaN	...	0.0	0.0	0.0	0.05	0.02	0.03	-0.03	-0.04	0.07	0.00
1	1	sighted	ripe-unripe	sighted_1	white	7	original	self	NaN	NaN	...	0.0	0.0	0.0	0.02	0.09	0.11	0.05	0.08	-0.17	0.09
2	2	sighted	new-old	sighted_1	white	1	original	self	NaN	NaN	...	0.0	0.0	0.0	0.12	0.04	0.16	0.07	0.08	0.06	0.05
3	3	sighted	submissive-aggressive	sighted_1	white	1	original	self	NaN	NaN	...	0.0	0.0	0.0	-0.08	-0.08	-0.05	-0.07	-0.06	-0.06	-0.07
4	4	sighted	selfless-jealous	sighted_1	white	1	original	self	NaN	NaN	...	0.0	0.0	0.0	0.05	-0.01	-0.00	-0.04	0.03	0.04	0.11
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
42975	28507	sighted	light-heavy	sighted_68129	red	5	replication_2	other	11.0	0.0	...	0.0	0.0	0.0	-0.16	-0.04	-0.07	-0.10	-0.14	-0.13	-0.05
42976	28508	sighted	relaxed-tense	sighted_68129	red	5	replication_2	other	11.0	0.0	...	0.0	0.0	0.0	0.11	-0.04	0.08	-0.03	-0.01	0.01	0.01
42977	28509	sighted	alive-dead	sighted_68129	red	6	replication_2	other	11.0	0.0	...	-1.0	0.0	0.0	0.15	0.06	0.04	0.07	-0.03	0.07	0.06
42978	28510	sighted	fast-slow	sighted_68129	red	3	replication_2	other	11.0	0.0	...	0.0	0.0	0.0	0.00	-0.02	-0.06	-0.01	-0.06	0.01	-0.02
42979	28511	sighted	high-low	sighted_68129	red	2	replication_2	other	11.0	0.0	...	0.0	0.0	0.0	-0.01	0.04	-0.03	-0.01	0.03	0.06	0.04

42980 rows × 33 columns

Filtered COCA

COCA embeddings, but from COCA corpora without sentences with 1st order cooccurrences (sentences with a color word and a dimension word).

def get_cosine(x, vecs_dict):
    zero = np.zeros(300)
    return np.dot(norm(vecs_dict.get(x['word1'], zero) - vecs_dict.get(x['word2'], zero)), vecs_dict.get(x['color'], zero))

# fiction
vecs = Vectors('../embeddings/fic.filtered.en.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()
df_joint['cosine_fic_filtered'] = df_joint.apply(lambda x: get_cosine(x, vecs_dict), axis=1)

[INFO] loading vectors ../embeddings/fic.filtered.en.vec
[INFO] <function Vectors.__init__ at 0x7fb0731c9d40> ran in 16.957 seconds
[INFO] <function Vectors.as_dict at 0x7fb0731c9f80> ran in 0.138 seconds

No neighbors COCA

COCA embeddings, but from training corpora from which the 100 nearest neighbors of each dimension word have been removed (in an attempt to disrupt the "scaffolding" that semantic associations with the dimension words are built on).

# fiction
vecs = Vectors('../embeddings/fic.noneighbors.en.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()
df_joint['cosine_fic_noneighbors'] = df_joint.apply(lambda x: get_cosine(x, vecs_dict), axis=1).fillna(0)

[INFO] loading vectors ../embeddings/fic.noneighbors.en.vec
[INFO] <function Vectors.__init__ at 0x7fb0731c9d40> ran in 15.297 seconds
[INFO] <function Vectors.as_dict at 0x7fb0731c9f80> ran in 0.130 seconds

No names COCA

COCA embeddings, but from training corpora from which the labels generated by at least two participants for color-semantic associations (e.g. the label snow for the combination white and cold) has been removed. (These nameability data are explored in more detail in a section at the end of this notebook.)

# fiction
vecs = Vectors('../embeddings/fic.nonames.en.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()
df_joint['cosine_fic_nonames'] = df_joint.apply(lambda x: get_cosine(x, vecs_dict), axis=1).fillna(0)

[INFO] loading vectors ../embeddings/fic.nonames.en.vec
[INFO] <function Vectors.__init__ at 0x7fb0731c9d40> ran in 14.704 seconds
[INFO] <function Vectors.as_dict at 0x7fb0731c9f80> ran in 0.139 seconds

Correlations between predictors

Correlations between predictors in original data (NZ)

df_orig = df_joint[df_joint['experiment'] == 'original']

corrs = np.abs(df_joint[[
    'rating',
    'cosine_cc',
    'cosine_subs',
    'cosine_fic',
    'cosine_fic_filtered',
    'cosine_fic_noneighbors',
    'cosine_fic_nonames',
    'swow_all',
    'swow_all_NZ',
    'swow_all_US',
    'frequency',
    'concreteness',
]].corr()).round(2)
g = sns.heatmap(corrs, vmin=0, vmax=1, annot=True)

Correlations between predictors in 2nd replication data (US)

df_rep = df_joint[df_joint['experiment'] == 'replication_2']

corrs = np.abs(df_joint[[
    'rating',
    'cosine_cc',
    'cosine_subs',
    'cosine_fic',
    'cosine_fic_filtered',
    'cosine_fic_noneighbors',
    'cosine_fic_nonames',
    'swow_all',
    'swow_all_NZ',
    'swow_all_US',
    'frequency',
    'concreteness',
]].corr()).round(2)
g = sns.heatmap(corrs, vmin=0, vmax=1, annot=True)

Correlations between predictors in full dataset

corrs = np.abs(df_joint[[
    'rating',
    'cosine_cc',
    'cosine_subs',
    'cosine_fic',
    'cosine_fic_filtered',
    'cosine_fic_noneighbors',
    'cosine_fic_nonames',
    'swow_all',
    'swow_all_NZ',
    'swow_all_US',
    'frequency',
    'concreteness',
]].corr()).round(2)
g = sns.heatmap(corrs, vmin=0, vmax=1, annot=True)

Standardize predictors and write to file

def standardize(Series):
    return (Series - Series.mean()) / Series.std()

df_joint['art_z'] = standardize(df_joint['art'])
df_joint['fiction_z'] = standardize(df_joint['fiction'])
df_joint['nonfiction_z'] = standardize(df_joint['nonfiction'])
df_joint['reading_motivation_z'] = standardize(df_joint['reading_motivation'])
df_joint['reading_part_of_self_z'] = standardize(df_joint['reading_part_of_self'])
df_joint['reading_efficacy_z'] = standardize(df_joint['reading_efficacy'])
df_joint['reading_recognition_z'] = standardize(df_joint['reading_recognition'])
df_joint['reading_other_realms_z'] = standardize(df_joint['reading_other_realms'])

df_joint['rating_z'] = standardize(df_joint['rating'])
df_joint['frequency_z'] = standardize(df_joint['frequency'])
df_joint['concreteness_z'] = standardize(df_joint['concreteness'])

df_joint['swow_all_z'] = standardize(df_joint['swow_all'])
df_joint['swow_all_NZ_z'] = standardize(df_joint['swow_all_NZ'])
df_joint['swow_all_US_z'] = standardize(df_joint['swow_all_US'])
df_joint['swow_R1_z'] = standardize(df_joint['swow_R1'])
df_joint['swow_R1_NZ_z'] = standardize(df_joint['swow_R1_NZ'])
df_joint['swow_R1_US_z'] = standardize(df_joint['swow_R1_US'])

df_joint['cosine_cc_z'] = standardize(df_joint['cosine_cc'])
df_joint['cosine_subs_z'] = standardize(df_joint['cosine_subs'])

df_joint['cosine_acad_z'] = standardize(df_joint['cosine_acad'])
df_joint['cosine_fic_z'] = standardize(df_joint['cosine_fic'])
df_joint['cosine_mag_z'] = standardize(df_joint['cosine_mag'])
df_joint['cosine_news_z'] = standardize(df_joint['cosine_news'])
df_joint['cosine_spok_z'] = standardize(df_joint['cosine_spok'])

df_joint['cosine_fic_filtered_z'] = standardize(df_joint['cosine_fic_filtered'])
df_joint['cosine_fic_noneighbors_z'] = standardize(df_joint['cosine_fic_noneighbors'])
df_joint['cosine_fic_nonames_z'] = standardize(df_joint['cosine_fic_nonames'])

df_joint['blind'] = pd.get_dummies(df_joint['group'])['blind']
df_joint['sighted'] = pd.get_dummies(df_joint['group'])['sighted']
df_joint['group_eff'] = (df_joint['sighted'] - .5) * 2
df_joint['group_z'] = standardize(df_joint['sighted'])

df_joint['original'] = pd.get_dummies(df_joint['experiment'])['original']
df_joint['replication_1'] = pd.get_dummies(df_joint['experiment'])['replication_1']
df_joint['replication_2'] = pd.get_dummies(df_joint['experiment'])['replication_2']

df_joint['other'] = pd.get_dummies(df_joint['self_vs_other'])['other']
df_joint['self'] = pd.get_dummies(df_joint['self_vs_other'])['self']
df_joint['self_vs_other_eff'] = (df_joint['other'] - .5) * 2
df_joint['self_vs_other_z'] = standardize(df_joint['other'])

df_joint.to_csv('data_plus_predictors.tsv', sep='\t', index=False)

display(df_joint)

	index	group	dimension	pp_id	color	rating	experiment	self_vs_other	art	fiction	...	sighted	group_eff	group_z	original	replication_1	replication_2	other	self	self_vs_other_eff	self_vs_other_z
0	0	sighted	cold-hot	sighted_1	white	1	original	self	NaN	NaN	...	1	1.0	0.211241	1	0	0	0	1	-1.0	-0.891882
1	1	sighted	ripe-unripe	sighted_1	white	7	original	self	NaN	NaN	...	1	1.0	0.211241	1	0	0	0	1	-1.0	-0.891882
2	2	sighted	new-old	sighted_1	white	1	original	self	NaN	NaN	...	1	1.0	0.211241	1	0	0	0	1	-1.0	-0.891882
3	3	sighted	submissive-aggressive	sighted_1	white	1	original	self	NaN	NaN	...	1	1.0	0.211241	1	0	0	0	1	-1.0	-0.891882
4	4	sighted	selfless-jealous	sighted_1	white	1	original	self	NaN	NaN	...	1	1.0	0.211241	1	0	0	0	1	-1.0	-0.891882
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
42975	28507	sighted	light-heavy	sighted_68129	red	5	replication_2	other	11.0	0.0	...	1	1.0	0.211241	0	0	1	1	0	1.0	1.121199
42976	28508	sighted	relaxed-tense	sighted_68129	red	5	replication_2	other	11.0	0.0	...	1	1.0	0.211241	0	0	1	1	0	1.0	1.121199
42977	28509	sighted	alive-dead	sighted_68129	red	6	replication_2	other	11.0	0.0	...	1	1.0	0.211241	0	0	1	1	0	1.0	1.121199
42978	28510	sighted	fast-slow	sighted_68129	red	3	replication_2	other	11.0	0.0	...	1	1.0	0.211241	0	0	1	1	0	1.0	1.121199
42979	28511	sighted	high-low	sighted_68129	red	2	replication_2	other	11.0	0.0	...	1	1.0	0.211241	0	0	1	1	0	1.0	1.121199

42980 rows × 74 columns

df_joint = pd.read_csv('data/data_plus_predictors.tsv', sep='\t')

Nameability of color-dimension associations

def get_cosine_1word(x, vecs_dict):
    zero = np.zeros(300)
    return np.dot(vecs_dict.get(x['dimension'], zero), vecs_dict.get(x['color'], zero))

# fiction
vecs = Vectors('../embeddings/fic.en.vec', n=2e5, d=300, normalize=True)  # not included in git repo
vecs_dict = vecs.as_dict()

df_names = pd.read_csv('data/color_dimension_nameability.csv')
display(df_names.head())

[INFO] loading vectors ../embeddings/fic.en.vec
[INFO] <function Vectors.__init__ at 0x13d7ac0d0> ran in 6.203 seconds
[INFO] <function Vectors.as_dict at 0x13d7ac310> ran in 0.056 seconds

	prompt	dimension	color	number_responses	avg_words_per_response	percent_unique_words	percent_unique_lemmas	simpson_diversity	modal_agreement	modal_names	modal_response_agreement	modal_response
0	happy_brown	happy	brown	10	1.000000	0.800000	0.800000	0.044444	0.200000	cat,puppy	0.200000	cat,puppy
1	unripe_brown	unripe	brown	10	1.000000	1.000000	1.000000	0.000000	0.100000	grape,bannana,kiwi,avocado,pear,fruit,tree,coc...	0.100000	grape,bannana,kiwi,avocado,pear,fruit,tree,coc...
2	hard_brown	hard	brown	10	1.000000	0.900000	0.800000	0.044444	0.200000	wood,rock	0.200000	wood
3	angry_blue	angry	blue	13	1.076923	0.714286	0.714286	0.054945	0.230769	shark	0.230769	shark
4	sad_brown	sad	brown	10	1.100000	0.909091	0.909091	0.018182	0.200000	cat	0.200000	cat

# check how many participants provided labels for each color-adjective pair
print(df_names['number_responses'].min())
print(df_names['number_responses'].max())
display(df_names.sort_values('modal_agreement'))

7
13

	prompt	dimension	color	number_responses	avg_words_per_response	percent_unique_words	percent_unique_lemmas	simpson_diversity	modal_agreement	modal_names	modal_response_agreement	modal_response
99	liked_blue	liked	blue	13	1.000000	1.000000	1.000000	0.000000	0.076923	sonic,sky,bird,pigeon,phone,smurfs,pencil,colo...	0.076923	sonic,sky,bird,pigeon,phone,smurfs,pencil,colo...
40	relaxed_blue	relaxed	blue	13	1.076923	1.000000	1.000000	0.000000	0.076923	smurfette,meditation,bird,water,tranquility,st...	0.076923	smurfette,meditation,bird,water,tranquility,st...
30	submissive_blue	submissive	blue	13	1.000000	1.000000	1.000000	0.000000	0.076923	macaw,nun,bird,swallow,butterfly,flower,door,b...	0.076923	macaw,nun,bird,swallow,butterfly,flowers,door,...
91	old_blue	old	blue	13	1.076923	1.000000	1.000000	0.000000	0.076923	bluecheese,necklace,bird,dress,shoe,smurfs,rug...	0.076923	bluecheese,necklace,bird,dress,shoes,smurfs,ru...
192	clean_yellow	clean	yellow	12	1.083333	1.000000	1.000000	0.000000	0.083333	table,detergant,sun,glove,hat,flag,ford,mustan...	0.083333	table,detergant,sun,gloves,hat,flag,ford.musta...
...	...	...	...	...	...	...	...	...	...	...	...	...
151	clean_white	clean	white	9	1.111111	0.700000	0.600000	0.222222	0.555556	sheet	0.333333	sheets
170	ripe_yellow	ripe	yellow	12	1.000000	0.500000	0.500000	0.318182	0.583333	banana	0.583333	banana
68	cold_blue	cold	blue	13	1.000000	0.461538	0.461538	0.358974	0.615385	ice	0.615385	ice
147	cold_white	cold	white	9	1.000000	0.333333	0.333333	0.583333	0.777778	snow	0.777778	snow
158	stale_white	stale	white	9	1.000000	0.222222	0.222222	0.777778	0.888889	bread	0.888889	bread

306 rows × 12 columns

Exporting names generated by participants for use in training corpus filtering

names = df_names['modal_names']
names = list(chain(*[name.split(',') for name in names]))
names_all = set(names)  # all unique names
names_count = Counter(names)
names_2plus = [name[0] for name in names_count.most_common() if name[1] >= 2]  # all names that occur 2+ times
print(f'Number of labels named by at least 2 participants: {len(names_2plus)}')
with open('data/pair_labels_all.txt', 'w') as namesfile:
    namesfile.write('\n'.join(names_all))
with open('data/pair_labels_2plus.txt', 'w') as namesfile:
    namesfile.writelines('\n'.join(names_2plus))
# let's ignore words like "me", "my", and "a" though

Number of labels named by at least 2 participants: 242

Correlating COCA-fiction cosine similarities to nameability measures

Since we only have nameability for colors and dimension axis poles (i.e. for yellow and dislike but not yellow and dislike-like), we correlate nameability measures with cosine similarity between color and dimension axis pole.

pearsonr(df_names['simpson_diversity'], df_names['modal_agreement'])

PearsonRResult(statistic=0.8947743710654124, pvalue=1.816739746708339e-108)

df_names['cosine_fic'] = df_names.apply(lambda x: get_cosine_1word(x, vecs_dict), axis=1)
display(df_names.head())

x = pearsonr(df_names['cosine_fic'], df_names['simpson_diversity'])
print(f'pearsonr(cosine_fiction, simpson_diversity): {x[0]:.3f}, p-value: {x[1]:.3f}')
x = pearsonr(df_names['cosine_fic'], df_names['modal_agreement'])
print(f'pearsonr(cosine_fiction, modal_agreement): {x[0]:.3f}, p-value: {x[1]:.3f}')
g = sns.lmplot(x='cosine_fic', y='simpson_diversity', data=df_names)
g = sns.lmplot(x='cosine_fic', y='modal_agreement', data=df_names)

	prompt	dimension	color	number_responses	avg_words_per_response	percent_unique_words	percent_unique_lemmas	simpson_diversity	modal_agreement	modal_names	modal_response_agreement	modal_response	cosine_fic
0	happy_brown	happy	brown	10	1.000000	0.800000	0.800000	0.044444	0.200000	cat,puppy	0.200000	cat,puppy	0.105870
1	unripe_brown	unripe	brown	10	1.000000	1.000000	1.000000	0.000000	0.100000	grape,bannana,kiwi,avocado,pear,fruit,tree,coc...	0.100000	grape,bannana,kiwi,avocado,pear,fruit,tree,coc...	0.322344
2	hard_brown	hard	brown	10	1.000000	0.900000	0.800000	0.044444	0.200000	wood,rock	0.200000	wood	0.211326
3	angry_blue	angry	blue	13	1.076923	0.714286	0.714286	0.054945	0.230769	shark	0.230769	shark	0.174112
4	sad_brown	sad	brown	10	1.100000	0.909091	0.909091	0.018182	0.200000	cat	0.200000	cat	0.228968

pearsonr(cosine_fiction, simpson_diversity): 0.214, p-value: 0.000
pearsonr(cosine_fiction, modal_agreement): 0.202, p-value: 0.000

Correlating group-averaged human ratings to nameability measure differentials.

Since we do not have human ratings for the association between colors and dimension axis poles (only for association between colors and dimension axes), we need to collapse our nameability measures for the two poles of each dimension axis. One way to do this is to compute difference scores.

df_sighted = df_joint.loc[df_joint['group'] == 'sighted']
df_sighted['diversity_word1'] = df_sighted.merge(df_names, how='left', left_on=['word1', 'color'], right_on=['dimension', 'color'])['simpson_diversity']
df_sighted['diversity_word2'] = df_sighted.merge(df_names, how='left', left_on=['word2', 'color'], right_on=['dimension', 'color'])['simpson_diversity']
df_sighted['agreement_word1'] = df_sighted.merge(df_names, how='left', left_on=['word1', 'color'], right_on=['dimension', 'color'])['modal_agreement']
df_sighted['agreement_word2'] = df_sighted.merge(df_names, how='left', left_on=['word2', 'color'], right_on=['dimension', 'color'])['modal_agreement']
df_sighted['diff_diversity'] = (df_sighted['diversity_word1'] - df_sighted['diversity_word2'])
df_sighted['diff_agreement'] = (df_sighted['agreement_word1'] - df_sighted['agreement_word2'])

df_sighted = df_sighted.dropna()
display(df_sighted.head())
df_mean_sighted = df_sighted.groupby(['color', 'dimension', 'word1', 'word2']).mean().reset_index()
df_sd_sighted = df_sighted.groupby(['color', 'dimension', 'word1', 'word2']).std().reset_index()
x = pearsonr(df_mean_sighted['rating'], df_mean_sighted['diff_diversity'])
print(f'pearsonr(rating, simpson_diversity_difference): {x[0]:.3f}, p-value: {x[1]:.3f}')
x = pearsonr(df_mean_sighted['rating'], df_mean_sighted['diff_agreement'])
print(f'pearsonr(rating, modal_agreement_difference): {x[0]:.3f}, p-value: {x[1]:.3f}')
x = pearsonr(df_mean_sighted['cosine_fic'], df_mean_sighted['diff_diversity'])
print(f'pearsonr(cosine_fiction, simpson_diversity_difference): {x[0]:.3f}, p-value: {x[1]:.3f}')
x = pearsonr(df_mean_sighted['cosine_fic'], df_mean_sighted['diff_agreement'])
print(f'pearsonr(cosine_fiction, modal_agreement_difference): {x[0]:.3f}, p-value: {x[1]:.3f}')

	index	group	dimension	pp_id	color	rating	experiment	self_vs_other	art	fiction	...	other	self	self_vs_other_eff	self_vs_other_z	diversity_word1	diversity_word2	agreement_word1	agreement_word2	diff_diversity	diff_agreement
14468	0	sighted	cold-hot	sighted_69212	brown	4	replication_2	self	3.0	0.0	...	0	1	-1.0	-0.891882	0.000000	0.012821	0.142857	0.285714	-0.012821	-0.142857
14469	1	sighted	ripe-unripe	sighted_69212	brown	7	replication_2	self	3.0	0.0	...	0	1	-1.0	-0.891882	0.238095	0.035714	0.428571	0.285714	0.202381	0.142857
14470	2	sighted	new-old	sighted_69212	brown	6	replication_2	self	3.0	0.0	...	0	1	-1.0	-0.891882	0.000000	0.000000	0.142857	0.142857	0.000000	0.000000
14471	3	sighted	submissive-aggressive	sighted_69212	brown	2	replication_2	self	3.0	0.0	...	0	1	-1.0	-0.891882	0.000000	0.000000	0.142857	0.142857	0.000000	0.000000
14472	4	sighted	selfless-jealous	sighted_69212	brown	5	replication_2	self	3.0	0.0	...	0	1	-1.0	-0.891882	0.000000	0.044444	0.142857	0.285714	-0.044444	-0.142857

5 rows × 80 columns

pearsonr(rating, simpson_diversity_difference): 0.036, p-value: 0.666
pearsonr(rating, modal_agreement_difference): -0.012, p-value: 0.890
pearsonr(cosine_fiction, simpson_diversity_difference): 0.058, p-value: 0.492
pearsonr(cosine_fiction, modal_agreement_difference): 0.059, p-value: 0.482

df_mean_sighted['rating_sd'] = df_sd_sighted['rating']
g = sns.lmplot(x='rating_sd', y='diff_diversity', data=df_mean_sighted)
x = pearsonr(df_mean_sighted['rating_sd'], df_mean_sighted['diff_diversity'])
print(f'pearsonr(rating_sd, simpson_diversity): {x[0]:.3f}, p-value: {x[1]:.3f}')
x = pearsonr(df_mean_sighted['rating_sd'], df_mean_sighted['diff_diversity'])
print(f'pearsonr(rating_sd, simpson_diversity): {x[0]:.3f}, p-value: {x[1]:.3f}')

pearsonr(rating_sd, simpson_diversity): 0.280, p-value: 0.001
pearsonr(rating_sd, simpson_diversity): 0.280, p-value: 0.001

Correlation of group-averaged split-inverse ratings with nameability measures

One other way to work around the issue of having only color to dimension axis pole nameability is to split and invert the human ratings of color-dimension axis associations to create two scores per rating: One for the right end of the axis (equal to the rating), and one for the left end of the axis (equal to eight minus the rating). For example: If yellow is assigned a 6 on the scale dislike-like, the rating for yellow/like is 6, but we also create a rating of 2 for yellow/dislike.

df_inverse = df_sighted[[
    'color',
    'word1',
    'rating',
    'diversity_word1',
    'agreement_word1'
]].rename(columns={
    'word1': 'dimension',
    'diversity_word1': 'simpson_diversity',
    'agreement_word1': 'modal_agreement'
})

df_inverse['rating'] = 8 - df_inverse['rating']

df_inverse = pd.concat([df_inverse, df_sighted[[
'color',
    'word2',
    'rating',
    'diversity_word2',
    'agreement_word2'
]].rename(columns={
    'word2': 'dimension',
    'diversity_word2': 'simpson_diversity',
    'agreement_word2': 'modal_agreement'
})])
display(df_inverse)

df_mean_inverse = df_inverse.groupby(['color', 'dimension']).mean().reset_index()
df_sd_inverse = df_inverse.groupby(['color', 'dimension']).std().reset_index()

x = pearsonr(df_mean_inverse['rating'], df_mean_inverse['simpson_diversity'])
print(f'pearsonr(rating, simpson_diversity): {x[0]:.3f}, p-value: {x[1]:.3f}')
x = pearsonr(df_mean_inverse['rating'], df_mean_inverse['modal_agreement'])
print(f'pearsonr(rating, modal_agreement): {x[0]:.3f}, p-value: {x[1]:.3f}')

	color	dimension	rating	simpson_diversity	modal_agreement
14468	brown	cold	4	0.000000	0.142857
14469	brown	ripe	1	0.238095	0.428571
14470	brown	new	2	0.000000	0.142857
14471	brown	submissive	6	0.000000	0.142857
14472	brown	selfless	3	0.000000	0.142857
...	...	...	...	...	...
41138	yellow	hard	2	0.000000	0.125000
41139	yellow	heavy	2	0.000000	0.125000
41140	yellow	tense	2	0.000000	0.125000
41141	yellow	dead	2	0.000000	0.125000
41142	yellow	slow	2	0.000000	0.125000

46272 rows × 5 columns

pearsonr(rating, simpson_diversity): 0.062, p-value: 0.293
pearsonr(rating, modal_agreement): 0.070, p-value: 0.237

df_mean_inverse['rating_sd'] = df_sd_inverse['rating']
g = sns.lmplot(x='rating_sd', y='modal_agreement', data=df_mean_inverse)
g = sns.lmplot(x='rating_sd', y='simpson_diversity', data=df_mean_inverse)
x = pearsonr(df_mean_inverse['rating_sd'], df_mean_inverse['simpson_diversity'])
print(f'pearsonr(rating_sd, simpson_diversity): {x[0]:.3f}, p-value: {x[1]:.3f}')
x = pearsonr(df_mean_inverse['rating_sd'], df_mean_inverse['modal_agreement'])
print(f'pearsonr(rating_sd, modal_agreement): {x[0]:.3f}, p-value: {x[1]:.3f}')

pearsonr(rating_sd, simpson_diversity): 0.228, p-value: 0.000
pearsonr(rating_sd, modal_agreement): 0.228, p-value: 0.000

In short: nameability (measured as simpson diversity and name agreement for the modal name) is weakly correlated with cosine similarity between colors and dimension axis poles, but not with human ratings, regardless of whether we fit the nameability to the ratings (by computing difference scores for the nameability measures) or fit the ratings to the nameability (by computing inverse ratings for the left poles of the dimension axes).

Extracting non-color nearest neighbors for each dimension

# reload COCA-fiction vecs
vecs = Vectors('../embeddings/fic.filtered.en.vec', n=1e6, d=300, normalize=True)
vecs_dict = vecs.as_dict()

color_vecs = filter_vecs(vecs, np.array(colors))
dimension_vecs = filter_vecs(vecs, np.array(dimensions))
dimension_pair_vecs = np.vstack([vecs_dict[pair[0]] - vecs_dict[pair[1]] for pair in dimension_pairs])

[INFO] loading vectors ../embeddings/fic.filtered.en.vec
[INFO] <function Vectors.__init__ at 0x7fb0731c9d40> ran in 17.975 seconds
[INFO] <function Vectors.as_dict at 0x7fb0731c9f80> ran in 0.153 seconds

Filtered 232340 vectors, 9 remaining.

[INFO] <function filter_vecs at 0x7fb0734b87a0> ran in 0.616 seconds

Filtered 232340 vectors, 34 remaining.

[INFO] <function filter_vecs at 0x7fb0734b87a0> ran in 0.801 seconds

dimension_neighbors = compute_nn(vecs, dimension_pair_vecs, np.array(dimension_labels), num_neighbors=100, whole_matrix=True)
dimension_neighbors = dimension_neighbors.rename(columns={'target': 'dimension'})
display(dimension_neighbors)
dimension_neighbors.to_csv('100_neighbors_coca_fic.tsv', sep='\t', index=False)

[INFO] <function Vectors.as_dict at 0x7fb0731c9f80> ran in 0.145 seconds
[INFO] computing analogies using whole matrix additive method
[INFO] <function compute_nn at 0x7fb0731c9cb0> ran in 0.951 seconds

	dimension	neighbor 0	neighbor 1	neighbor 2	neighbor 3	neighbor 4	neighbor 5	neighbor 6	neighbor 7	neighbor 8	...	neighbor -10	neighbor -9	neighbor -8	neighbor -7	neighbor -6	neighbor -5	neighbor -4	neighbor -3	neighbor -2	neighbor -1
0	cold-hot	cold	coldness	chill	icy	chilling	icy_cold	chilled	colder	chilly	...	burnin	Bellissimo	spics	hotting	hotspots	hottest	hotter	hotspot	Hot	hot
1	ripe-unripe	ripe	Forties	Belmont	Intimidating	riper	Corrupt	Proxy	high_maintenance	Facilities	...	kopf	tablets	uncap	unread	hoar	unsay	unrisen	unre	uns	unripe
2	new-old	new	fresh	new_batch	changes	rethink	redefining	sequencing	newly	newest	...	old_duffer	oldtimer	ancient	gnarled	Old	older	old_hag	old_farts	old_fart	old
3	submissive-aggressive	submissively	submissive	submission	submissions	submissiveness	Submission	Daughter_Lord	submit	submits	...	aggressor	aggressiveness	aggro	Aggressive	aggression	unaggressive	aggressions	aggres	aggressively	aggressive
4	selfless-jealous	selfless	selflessly	selflessness	selfless_act	ethos	wolfless_men	self_regard	self_denial	selfserving	...	snoop_around	jealous_rage	upset	jealousies	Jealous	jealous_type	jealousy	jeal	jealously	jealous
5	active-passive	active	activ	active_duty	actively	health_clubs	Glacier_Bay	Inactive	fund_raising_event	fund_raising	...	Passive	impassively	graceless	impassivity	submissiveness	passivity	impassive	Impassive	passively	passive
6	like-dislike	like	Like	Titty_Twister	Feels_like	raggedy_ass	Stelly	horror_movie	papier_mch	B_movie	...	broad_mindedness	antagonism	an_instant_dislike	intense_dislike	Dislike	mislike	disliking	disliked	dislikes	dislike
7	clean-dirty	clean	cleaned	cleancut	sterilize	reasonably	sterilized	cleanser	cleanses	cleans	...	dirty_slush	muddy	scurrilous	dirty_jokes	grimy	filthy	grubby	irty	Dirty	dirty
8	fresh-stale	fresh	afresh	freshly	Fresh	newly	new	clean	freshly_cut	refresh	...	stale_odor	stale_tobacco	stalemated	stale_cigarettes	staleness	stale_smoke	stale_air	stale_beer	stale_cigarette_smoke	stale
9	calm-angry	calm	calming	serene	calms	cool	some_semblance	Breathe_deep	soothing	calmness	...	humiliated	enraged	disgusted	furious	unhappy	angry_hornets	angrier	angry_mob	Angry	angry
10	happy-sad	happy	happ	blissfully_happy	perfectly_happy	oblige	supremely_happy	delighted	deliriously_happy	thrilled	...	woebegone	forlorn	sadness	sorrowful	wifeless	sadder	lifeless	mournfulness	mournful	sad
11	exciting-dull	exciting	Exciting	excited	excitingly	most_exciting	interesting	unexciting	excite	terribly_exciting	...	dulling	dull_throbbing	duller	dull_sheen	dull_eyed	dully	dulled	dull_dull	dulls	dull
12	soft-hard	soft	soft_murmur	soft_moan	purr	rustle	mewing	throaty	whispery	silky	...	dif_ficult	hardest	Hard	Difficult	difficult	centrate	harder_than	proving_difficult	harder	hard
13	light-heavy	light	illumination	dawnlight	lights	sunlight	wan_light	halogen_lights	illumine	illumi	...	heavily_laden	heavy_wool	eavy	heavies	hefty	heavy_load	laden	thick	Heavy	heavy
14	relaxed-tense	relaxed	relaxer	relaxing	relax	Relaxed	eased	Contented	muscles_relaxed	contented	...	final_confrontation	desperate_straits	Intense	Tense	nervous	accusatory	fearful	tenser	tenses	tense
15	alive-dead	alive	pleasurably	58_9	stay_alive	aloof	under_control	hardly_contain	marvelously	could_hardly_contain	...	empty_beer_cans	muddy_ditch	fallen	Dead	graves	laundryman	grave	Dropped	deader	dead
16	fast-slow	fast	quickly	Quickly	Ouickly	rapidly	faster_faster	Faster_faster	soon	easily	...	mesmerizing	deliberate	cadenced	quiescence	unremitting_courtesy	slow_drawl	languor	languorous	gentle	slow

17 rows × 201 columns

More figures

Mean color ratings on each dimension

df_viz = df_joint[df_joint['dimension'] != 'high-low']

df_means = df_viz.groupby(['dimension', 'color', 'word1', 'word2']).mean().reset_index()
dim_order = df_means.groupby('dimension').std().sort_values('rating', ascending=False).reset_index()['dimension']
df_means = df_means.set_index('dimension').loc[dim_order].reset_index()

mins_idx = df_means.groupby(['dimension'])['rating'].transform(min) == df_means['rating']
mins = df_means[mins_idx]
maxs_idx = df_means.groupby(['dimension'])['rating'].transform(max) == df_means['rating']
maxs = df_means[maxs_idx]

df_mins = mins[['word2', 'dimension', 'color']].merge(df_viz[['word2', 'dimension', 'color', 'rating']], how='left', on=['dimension', 'color', 'word2'])
df_maxs = maxs[['word1', 'dimension', 'color']].merge(df_viz[['word1', 'dimension', 'color', 'rating']], how='left', on=['dimension', 'color', 'word1'])

display(df_mins)
display(df_maxs)

	word2	dimension	color	rating
0	hot	cold-hot	blue	1
1	hot	cold-hot	blue	1
2	hot	cold-hot	blue	1
3	hot	cold-hot	blue	1
4	hot	cold-hot	blue	2
...	...	...	...	...
4555	passive	active-passive	red	2
4556	passive	active-passive	red	5
4557	passive	active-passive	red	2
4558	passive	active-passive	red	3
4559	passive	active-passive	red	3

4560 rows × 4 columns

	word1	dimension	color	rating
0	cold	cold-hot	red	7
1	cold	cold-hot	red	7
2	cold	cold-hot	red	7
3	cold	cold-hot	red	6
4	cold	cold-hot	red	7
...	...	...	...	...
4547	active	active-passive	brown	6
4548	active	active-passive	brown	5
4549	active	active-passive	brown	4
4550	active	active-passive	brown	4
4551	active	active-passive	brown	3

4552 rows × 4 columns

sns.set_style('darkgrid')
all_colors = {color: color for color in df_viz['color']}

fig, ax1 = plt.subplots(figsize=(3, 8))
sns.pointplot(data=df_viz, y='word1', x='rating', hue='color',
               palette=all_colors, join=False, dodge=False, ax=ax1, errorbar=('ci', .95))

ax2 = ax1.twinx()
sns.pointplot(data=df_viz, y='word2', x='rating', hue='color',
               palette=all_colors, join=False, dodge=False, ax=ax2, errorbar=('ci', .95))

ax1.set(ylabel='')
ax2.set(ylabel='')
ax1.get_legend().remove()
ax2.get_legend().remove()
ax1.set(xlim=[1, 7], xticks=[1, 2, 3, 4, 5, 6, 7]);

sns.set_style('whitegrid')
mins_colors = {color: color for color in mins['color']}
maxs_colors = {color: color for color in maxs['color']}

fig, ax1 = plt.subplots(figsize=(3, 7))
sns.violinplot(data=df_maxs, y='word1', x='rating', hue='color', #scale='width',
               palette=maxs_colors, dodge=False, ax=ax1, inner=None, cut=0)

ax2 = ax1.twinx()
sns.violinplot(data=df_mins, y='word2', x='rating', hue='color', #scale='area',
               palette=mins_colors, dodge=False, ax=ax2, inner=None, cut=0)

plt.setp(ax1.collections, alpha=.8)
plt.setp(ax2.collections, alpha=.8)
ax1.set(ylabel='')
ax2.set(ylabel='')
ax1.get_legend().remove()
ax2.get_legend().remove()
ax1.set(xlim=[1, 7], xticks=[1, 2, 3, 4, 5, 6, 7])
plt.savefig('figures/color_ratings.pdf', bbox_inches='tight')

Scatterplot with connected points

sns.set_style('darkgrid')
df_blind = df_viz[df_viz['group'] == 'blind'].groupby(['group', 'dimension', 'color', 'word1', 'word2']).mean().reset_index()
df_sighted = df_viz[df_viz['group'] == 'sighted'].groupby(['group', 'dimension', 'color', 'word1', 'word2']).mean().reset_index()
df_scatter = pd.concat([df_blind, df_sighted])
df_scatter['colordim'] = df_scatter['color'] + '_' + df_scatter['dimension']
df_scatter = df_scatter.sort_values('cosine_fic_z')
means_colors = {row['color']: row['color'] for _, row in df_scatter.iterrows()}
g = sns.FacetGrid(df_scatter, hue='color', col='group',  height=5, palette=means_colors, aspect=.5, sharex=True)
g.map(plt.scatter, 'cosine_fic_z', 'rating', s=10)
g.map(sns.regplot, 'cosine_fic_z', 'rating', scatter=False, ci=False)#, linewidth=.5)

g.set(xlabel='COCA-fiction\nembedding projection')
g.axes[0][0].set(ylabel='mean participant rating')
g.axes[0][0].set(title='blind')
g.axes[0][1].set(title='sighted')
g.set(ylim=[.75, 7.25], xlim=[-2.9, 2.9])
plt.savefig('figures/scatter_color.pdf', bbox_inches='tight')

Convert notebook to html

convert_notebook('data_prep')

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-11-07966e3605ae> in <module>
----> 1 convert_notebook('data_prep')

<ipython-input-1-ebaa396ce086> in convert_notebook(title, output)
     33     convert = sp.run(f'jupyter nbconvert {title}.ipynb --to {output} --output {title}.{output}'.split(' '))
     34     if convert.returncode == 0:
---> 35         display_md(f'Jupyter notebook `{title}` converted successfully.')
     36     else:
     37         display_md(f'Error: encountered problem converting Jupyter notebook `{title}`')

<ipython-input-1-ebaa396ce086> in display_md(md, **kwargs)
     28 
     29 def display_md(md, **kwargs):
---> 30     return display_markdown(md, raw=True, **kwargs)
     31 
     32 def convert_notebook(title, output='html'):

NameError: name 'display_markdown' is not defined